基于LDA话题演化研究方法综述_单斌.pdfVIP

24 6 Vol. 24, No . 6 2010 11 JOU RNAL OF CHINESE INFORMAT ION PROCESSING Nov., 2010 : 1003-00 ( 2010) 06-0043-0 LDA , ( , 200240) : 现实生活中不断有新话题的产生和旧话题的衰减, 同时话题的内容也会随着时间发生变化自动 测话 题随时间的演化越来越受到人们的关注Latent Dirichlet Allocation 模型是近年提出的概率话题模型, 已经在话 题演化领域得到较为广泛的应用该文提出了话题演化的两 个方面: 内容演化和强度演化, 总结了基于LDA 话题 模型的话题演化方法, 根据引入时间的不同方式将目前的研究方法分为三类: 将时间信息结合到LDA 模型对文 本集合后离散和先离散方法在详细叙述这三种方法的基础上, 针对时间粒度是否在线等多 个特征进行了对比, 并且简要描述了目前广泛应用的话题演化评测方法文章最后分析了目前存在的挑战, 并且对该研究方向进行了 展望 : 话题模型; 话题演化; Latent Dirichlet Allocation : T P391 : A A Survey of Topic Evolution Based on LDA SHAN Bin, LI Fang ( Sino-German Joint research Lab for Language Technologies, Dept. of Computer Science Engineering, Shanghai Jiao Tong University, Shanghai 200240, China) Abstract: With t opics evolve over time, new topics emerge and old ones decay . Many researches are devoted to de- tect the topic evolution automat ically. Latent Dirichlet Allocat ion ( LDA) , as a recently emerged probabilistic t opic model, has been w idely used in the research of topic evolut ion. This paper discusses two aspects of evolut ion on top- ic, i. e. the content and the topic intensit y. It summarizes three methods in LDA based topic evolution detection ac- cording to the dealing with time: joining the t ime to LDA model, post-discretizing or pre-discretizing methods. The three methods are also compared in several features: the time granularit y, on-line or off-line, etc. In addit ion, the e- valuation methods for topic evolution are int roduced. Finally, the paper gives some analysis and suggest ions for fu- ture researches on topic evolution based on LDA . Key words: t opic model; t opic evolut ion; Laten